Research On Intelligent Restoration Technology Of Ceramic Broken Cultural Relics Based On Point Cloud Completion

With the gradual integration of artificial intelligence and data mining into social life,more and more work is replaced by computers,which not only brings convenience to people’s life,but also saves costs and improves efficiency.The splicing recovery of ceramic fragments has always been a difficult problem in the field of cultural relics.The human and material costs of manual splicing recovery are large and inefficient.There are many difficulties in the research and implementation of 3D fragments intelligent recovery at home and abroad,such as the single implementation method,the low accuracy of matching and splicing fragments,the low accuracy of judging whether they belong to the same type,and the details of the fragments can not be completed well when the point cloud completion technology is used to complete the fragments.Based on point cloud completion,this thesis uses similarity measurement to realize intelligent restoration of ceramic broken cultural relics.The main work and innovation of this thesis are as follows:(1)Point cloud data preprocessingTo address the problem of disorganized point cloud data,the original data set is firstly preprocessed to facilitate the implementation of downstream tasks at a later stage.The process of data preprocessing mainly includes: taking multi-angle photos of 3D objects,removing background plates,keying operations,using 3D modeling software Photo Scan for 3D modeling of the photographed and processed 2D pictures,and getting the point cloud data of fragments,through which the modeled 3D objects are more realistic and practical than the 3D objects scanned by the scanner;the farthest point sampling method was used to achieve uniform downsampling of point cloud;the point cloud data is normalized with the center as the origin,and the point cloud coordinate value is scaled to the range of [-1,1].Through the preprocessing of point cloud data,the point cloud data is standardized,and the completion effect of the model is improved.(2)Point cloud completion method based on encoder and decoderIn this thesis,we propose a GAN-based Dynamic Point Fractal Network(DPF-Net),which can be used for accurate and high-fidelity point cloud completion to obtain the missing part of fragmentation completion.The generator in DPF-Net is composed of a multi-scale encoder and a multi-stage decoder.The encoder based on conditional convolution and attention mechanism extracts multi-scale features at different resolutions,which not only considers global features but also retains local details.The decoder completes the point cloud in stages with a fully connected backbone architecture,which not only completes the overall shape,but also completes the local details;The discriminator is used to optimize the encoder and decoder to improve the realism of the generated missing parts,and improve the phenomenon that the features of the same class can influence each other.The experimental results show that,compared with the current mainstream point cloud completion methods,the proposed method improves the accuracy and generalization ability of point cloud completion.(3)Similarity measurementIn order to improve the accuracy of fragment matching and recovery,and to accurately determine whether the fragments belong to the same ceramic type,two methods are used in this thesis for similarity measurement,namely,point cloud registration and similarity comparison of surface curvature.This method compares the similarity of the completion part of the original fragment and other fragments in two ways.Firstly,the similarity comparison is performed from the point perspective by applying the iterative nearest point registration method of point-topoint to determine the degree of similarity in terms of visualization results and quantitative metrics;secondly,from the surface similarity perspective,the surface curvatures of the feature points on the two fragments are calculated separately to form two surface curvature feature vectors,and then the Euclidean distance and Jaccard distance of the two feature vectors are used to determine the similarity between them.The Euclidean distance and Jaccard distance of the two eigenvectors are used to determine the similarity between the two fragments,that is,the final fragment that can be spliced is identified and combined with the original fragment to achieve recovery.Meanwhile,for the small fragments,the quantitative index of the iterative nearest point registration and surface curvature can be used to determine whether the two fragments belong to the same type.(4)Intelligent recovery of ceramic fragments based on point cloud completion technologyIn this thesis,the point cloud completion technique is directly applied to the intelligent splicing and recovery of ceramic fragments.Firstly,the missing part of the large original fragment is obtained by the point cloud completion method,and then the missing part of the completion is compared with the fragments in the real fragment set.Finally,the fragment that can be spliced with the original fragment in the real fragment set is found by similarity measurement.In addition,part of the missing fragments of the complete artifacts can be intellitively completed to obtain the visualization result of the complete artifacts.Using this technology,it can not only realize the accurate splicing of the real fragments,but also realize the virtual repair of the incomplete artifacts.The experiments show that the method is feasible and has strong practical value.